Investigating Modality Contribution in Audio LLMs for Music

Models may not be "listening" as much as we think: The study found that the higher-performing Audio LLM actually relied more on text than audio to answer questions about music, raising concerns about whether these systems are genuinely processing audio information or primarily using textual reasoning.

Low overall audio contribution doesn't mean audio is ignored: Despite finding that audio's overall contribution was relatively low, the researchers discovered that models could still successfully identify and localize specific sound events, suggesting the audio modality plays a more nuanced role than simple contribution scores might indicate.

Explainability tools are essential for multimodal AI: This research demonstrates the value of using Shapley value-based frameworks (MM-SHAP) to understand how different data types contribute to AI decisions, highlighting that we need better methods to audit what multimodal systems actually rely on.

Current benchmarks may not reveal the full picture: The paper suggests that standard accuracy metrics alone are insufficient for evaluating Audio LLMs—models can achieve good performance while potentially using modalities in unexpected ways, underscoring the need for deeper investigation into how these systems work.

Responsible deployment requires understanding model behavior: For organizations deploying Audio LLMs in real-world applications, this research emphasizes that knowing how a model reaches its conclusions is as important as knowing if it's accurate, particularly when users expect genuine audio understanding.